Add laduck: LLM inference inside DuckDB via llama.cpp

[hamidr]

Summary

Adds laduck, a DuckDB extension that embeds llama.cpp for local LLM inference directly from SQL.

• Text generation: llm_complete(text, model)
• Embeddings: llm_embed(text, model) returns float arrays
• Zero-shot classification: llm_classify(text, model, labels) in a single forward pass
• Model sources: Local files, HuggingFace (hf://), or DuckDB internal storage (db://)
• Portable databases: Store GGUF models inside .duckdb files
• GPU acceleration: Metal (macOS), CUDA, Vulkan, ROCm

Build notes

llama.cpp is included as a git submodule (third_party/llama.cpp) and statically linked. The submodule adds build time but has no runtime dependencies. WASM platforms are excluded since llama.cpp requires native execution.

Repository

https://github.com/hamidr/laduck